To maintain compatibility with current workflows while adding additional functionality to the workflow, we intend to change the way perturbed source parameters are passed to the components of the slurm ground motion workflow.
It has been proposed that srf generation will be split into two steps: perturbation and srf generation.
This will split the current two workflow steps (gcmt2srf, nhm2srf) to generate srfs into three (gcmt2srfparams, nhm2srfparams, srfparams2srf).
There will be an intermediary file containing all the parameters to generate the srf, allowing users to manually modify these files if they choose to do so.
This will be able to be integrated into the automated workflow by making each step work on a single realisation, with a multi process wrapper to allow it to also be used as a stand alone step. The automated workflow step would not allow time for the user to modify the intermediary file.
All parameters to be perturbed are to be included in the source_uncertainty.yaml file. Any additional parameters must also have nominal values provided.
The nominal values and perturbation settings for each of these parameters will have to be given in the source uncertainties yaml (as passed to gcmt2srf).
Each parameter that is given in the CMT solutions csv file will be at the top level, while each additional parameter will be nested inside the relevant process.
Each parameter will have at least two associated pieces of information, being the nominal value (mean) of the parameter, and the distribution. If the distribution requires any random variables, these must also be given.
An example is available below:
mw:
distribution: uniform
half_range: 0.05
hf:
sdrop:
mean: 50
half_range: 5
distribution: uniform
rvfac:
mean: 0.8
std_dev: 0.01
distribution: normal
rvfac_shallow:
mean: 0.7
distribution: none
bb:
fmin:
mean: 0.2
distribution: uniform_relative
scale_factor: 0.05
emod3d:
fmin:
mean: 0.01
std_dev: 0.001
distribution: normal |
| Distribution name | Distribution variable | Notes |
|---|---|---|
| none | N/A | Uses the value as given. |
| uniform | half_range | mean is the median (middle) value, and half_range half the difference between the highest and lowest value. |
| normal | std_dev | Regular normal distribution. |
| uniform_relative | scale_factor | A normal distribution where the highest and lowest possible value are mean* (1+scale_factor) and mean* (1-scale_factor) respectively. Scale factor to be given as a decimal. |
| log_normal | std_dev | Regular log_nromal distribution. |
| weibull | k, scale_factor | Does not take a mean value. Equivalent to scale_factor*(-ln((U))1/a where U is the uniform distribution over [0, 1). |
| truncated_normal | std_dev, std_dev_limit | Similar to normal distribution, except all values fall within [std_dev_limit] standard deviations of the mean. std_dev_limit has a default value of 2. |
| truncated_log_normal | std_dev, std_dev_limit | Similar to log normal distribution, except all values of the underlying normal distribution fall within [std_dev_limit] standard deviations of the mean. std_dev_limit has a default value of 2. Mean must be given as a string of the form "log(x)" where x is the unmodified value of the mean. |
| duplicate | targets (must be a list of strings, even if only 1 target) | Copies the given target perturbated variable. The perturbate_parameters function must include functionality to handle cases where the target is perturbated after the current variable. |
Any variable found in the e3d.par file may be placed in the emod3d section of source_uncertainty.yaml, and will be carried over. Any non numeric entries must have the distribution 'none'.
In order to simplify the implementation of this, we have chosen to have every parameter added to the slurm script and passed as an argument to the simulation script, which will then discard any parameters it does not recognize.
These parameters have a mean specified via other methods and do not need one in the source uncertainty file. If a mean is provided in the source uncertainty file it will take precedence
| Parameter | Type 1 | Type 2 | Type 3 | Type 4 |
|---|---|---|---|---|
| depth | ✓ | ✓ | ||
| mw (magnitude) | ✓ | ✓ | ✓ | ✓ (mag0) |
| mom | ✓ | ✓ (mom0) | ||
| strike | ✓ | ✓ | ✓ | ✓ (stk) |
| rake | ✓ | ✓ | ✓ | ✓ (rak) |
| dip | ✓ | ✓ | ✓ | ✓ |
| vs | ✓ | |||
| rho | ✓ | |||
| rise_time | ✓ | |||
| rvfac | Optional | Optional | Optional | |
| rough | Optional | Optional | Optional | |
| slip_cov | Optional | Optional | Optional | |
| flen | Calculated, but can be given/perturbated | ✓ | ✓ | |
| dlen | Calculated, but can be given/perturbated | ✓ | ✓ | |
| fwid | Calculated, but can be given/perturbated | ✓ | ✓ | |
| dwid | Calculated, but can be given/perturbated | ✓ | ✓ | |
| dtop | Calculated, but can be given/perturbated | ✓ | ✓ | |
| shypo | Calculated, but can be given/perturbated | ✓ | ✓ | |
| dhypo | Calculated, but can be given/perturbated | ✓ | ✓ |
Genslip takes the parameter rvfrac while HF takes the parameter rvfac, however scientifically these are the same value.
This different spelling allows the two to be perturbated independently. With the current perturbation implementation a challenge arises when it is to be perturbated and used for both.
The currently proposed solution is that:
The parameters for each component are as follows:
| hf | bb | |
|---|---|---|
| flo | ✓ | |
| fmin | ✓ | |
| fmidbot | ✓ | |
| lfvsref | ✓ | |
| sim_bin | ✓ | |
| t-sec | ✓ | |
| sdrop | ✓ | |
| rayset | ✓ | |
| seed | ✓ | |
| duration | ✓ | |
| dt | ✓ | |
| fmax | ✓ | |
| kappa | ✓ | |
| qfexp | ✓ | |
| rvfac | ✓ | |
| rvfac_shallow | ✓ | |
| rvfac_deep | ✓ | |
| czero | ✓ | |
| calpha | ✓ | |
| mom | ✓ | |
| rupv | ✓ | |
| velocity-model | ✓ | |
| site-vm-dir | ✓ | |
| vs-moho | ✓ | |
| fa_sig1 | ✓ | |
| fa_sig2 | ✓ | |
| rv_sig1 | ✓ | |
| hf_path_dur | ✓ |